On Wed, Nov 23, 2011 at 08:24, jean-frederic thebault < jean-frederic at thebault-net.com> wrote:
> I'm wondering what's wrong in my code. I'm using PETSc to solve a linear > system, and willing to use a multi-processor computer. 9 years ago, I used > petsc-2.1.3 with success. Few weeks ago, I've update petsc with the 3.1-p8 > version and made the necessary changes to work with. No problem. And > recently, I've migrate to petsc-3.2-p5. Compilation is OK. But when I do > simulation, now, I have some PETSC-ERROR in the log file, even using only > one processor (see the out.log file in this email). > You are calling MatSetOption() with the wrong number of arguments. C compilers tell you about this, but Fortran compilers do not. http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/Mat/MatSetOption.html > However, I think I defined MatMPI and VecMPI correctly, according to the > doc. The log file tell that something wrong with the nnz which should not > be greater than row length (??). > The log you sent does not say anything about nnz. Fix the call to MatSetOption(). And also, with the previous version of PETSc I've used, the were no problem > using -pc_type bjacobi and -sub_pc_type sor, juste to solve linear system > doing parallel computations and because SOR is not parallelized. But now, > when I use -pc_type bjacobi and -sub_pc_type sor, with 3 rank, I experiment > some convergence problem during my simulation. > These options should do the same thing they used to do. Make sure you are assembling correctly. If it's still confusing, run the old and new code with -ksp_monitor_true_residual -ksp_converged_reason -ksp_view -pc_type bjacobi -sub_pc_type sor and send the output of both for us to look at. Also note that you can use -pc_type sor even in parallel. There are options for local iterations and full iterations. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20111123/3d7ef281/attachment.htm>
